RESILIENCE IN HUMAN–AI–ROBOT TEAMS: WIDENING THE SCOPE OF MEASUREMENT

Project overview. As Human–AI–Robot Teams (HARTs) become prevalent in safety-critical domains, team resilience becomes increasingly relevant for assessing their effectiveness. This study explores a dynamical systems approach to connect interaction-based measures of nominal teamwork with processes and outcomes related to positive adaptation in perturbed team contexts. Method. An experiment was conducted using a remotely-accessible testbed based on Next Generation Combat Vehicle concepts. Groups of 3-participants were recruited to form teams and completed two ground reconnaissance missions set in Minecraft involving movement, communication, and the identification of battlefield objects such as infantry and obstacles. Throughout these missions, teams experienced novel perturbations. Two novel perturbations are the focus of this study: 1) the Lost Connection perturbation (LC) in Mission 1, which disabled one combat vehicle’s mobility, requiring repair from the other vehicle, and 2) the Lost Visibility perturbation (LV) in Mission 2, which restricted one combat vehicle’s vision to three blocks away (Minecraft units) during a task to transport a civilian to a helicopter extraction site. The measures reported here include measures of team performance in response to perturbations in overall relaxation time (Tvotzky et al., 2012) and broken down across problem solving components (Hoffman and Hancock 2017), and dynamical systems measures of communication turn-taking, determinism (Gorman et al., 2012) and the Hurst exponent (Eke et al., 2002). The previously described measures were applied to two teams, denoted as Team A and Team B. These teams were selected based on their good (Team A) and bad (Team B) responses to the LV perturbation. Preliminary results and discussion. Relaxation time was similar across teams for the LC perturbation, but was much shorter for Team A (RT = 278 seconds) than for Team B (RT = 841 seconds) for the LV perturbation. Team A implemented their solution to the LV perturbation faster (I = 8.99 than Team B (I = 15.34) relative to the overall event time. For Team A, the Hurst exponent indicated the determinism of turntaking series was anti-persistent for Mission 1, H = 0.389, as well as for Mission 2, H = 0.295. Surrogate analyses (Dolan and Spano, 2001) indicated that H for Team A were significantly different from randomly shuffled series, p < 0.05. The Hurst exponents for Team B suggested that the series for Team B were near random for both Mission 1, H = 0.476, and Mission 2, H = 0.426. An anti-persistent pattern of communication determinism indicates that short-term changes in a team's level of organization are inversely related to long-term changes. This was the case for Team A, but not Team B, which had near-random organization. These preliminary results suggest that resilient teams may have control over their organization throughout a mission, as it appears the team with worse adaptation to the LV perturbation had near random patterns of organization for both missions. Although measures of team resilience tend to focus on team processes or outcomes around perturbations, research and practice should also consider measuring teamwork in normal conditions to detect a team’s state of resilience. Developing continuous measures of team resilience may aid the development of Next Generation Combat Vehicle concepts and similar HART contexts.